ppbrown commited on
Commit
8a7d64c
1 Parent(s): 13f97c3

Upload T5/generate-dict-embeddingsT5.py with huggingface_hub

Browse files
Files changed (1) hide show
  1. T5/generate-dict-embeddingsT5.py +74 -0
T5/generate-dict-embeddingsT5.py ADDED
@@ -0,0 +1,74 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/bin/env python
2
+
3
+ """
4
+ (T5 counterpart of "generate-dict-embeddingsXL.py".
5
+ """
6
+
7
+ outputfile="embeddingsT5.temp.safetensors"
8
+
9
+ import sys
10
+ import torch
11
+ from safetensors.torch import save_file
12
+ from transformers import T5Tokenizer,T5EncoderModel
13
+
14
+ processor=None
15
+ tmodel=None
16
+
17
+ device=torch.device("cuda")
18
+
19
+ def initT5model():
20
+ global processor,tmodel
21
+ T="mcmonkey/google_t5-v1_1-xxl_encoderonly"
22
+ processor = T5Tokenizer.from_pretrained(T)
23
+ tmodel = T5EncoderModel.from_pretrained(T).to(device)
24
+
25
+
26
+ def embed_from_text(text):
27
+ global processor,tmodel
28
+ #print("Word:"+text)
29
+ tokens = processor(text, return_tensors="pt")
30
+ tokens.to(device)
31
+
32
+ if len(tokens.input_ids) >2:
33
+ print("ERROR: expected single token per word")
34
+ print(text)
35
+ exit(1)
36
+
37
+ with torch.no_grad():
38
+ outputs = tmodel(tokens.input_ids)
39
+
40
+ embedding = outputs.last_hidden_state[0][0]
41
+ #print(encoding.shape)
42
+ # Shape of this is (1,2,4096)
43
+
44
+ return embedding
45
+
46
+ initT5model()
47
+
48
+ print("Reading in 'dictionary'")
49
+ with open("dictionary","r") as f:
50
+ tokendict = f.readlines()
51
+ tokendict = [token.strip() for token in tokendict] # Remove trailing newlines
52
+
53
+
54
+ count=1
55
+ all_embeddings = []
56
+
57
+ for word in tokendict:
58
+ emb = embed_from_text(word)
59
+ #emb=emb.unsqueeze(0) # stupid matrix magic to make torch.cat work
60
+ all_embeddings.append(emb)
61
+ count+=1
62
+ if (count %100) ==0:
63
+ print(count)
64
+
65
+
66
+ embs = torch.cat(all_embeddings,dim=0)
67
+ print("Shape of result = ",embs.shape)
68
+
69
+ if len(embs.shape) != 2:
70
+ print("Sanity check: result is wrong shape: it wont work")
71
+
72
+ print(f"Saving the calculatiuons to {outputfile}...")
73
+ save_file({"embeddings": embs}, outputfile)
74
+